license: apache-2.0
base_model:
- Qwen/Qwen3.8-Flash-Next
base_model_relation: quantized
quantized_by: Synchopate
library_name: gguf
pipeline_tag: image-text-to-text
tags: - gguf
- qwen3.8
- qwen4exp
- moe
- abliterated
- uncensored
- 32gb-ram
- maxq
- cpu-inference
- llama-cpp
- llama-modchip
Qwen3.8-Flash-Next-MaxQ-32GB-Abliterated-GGUF (24.98 GiB Resident Trunk + 10.73 GiB Lazy NVMe PLE · Opt-In Subspace-Abliterated)
[!CAUTION]
Responsible-Use Disclaimer & Edition Positioning (Opt-In KL-Safe Subspace-Abliterated Checkpoint):Qwen3.8-Flash-Next-MaxQ-32GB-Abliteratedattenuates refusal-mediating activation subspaces via norm-preserving rank-2 orthogonal projection strictly acrossblk.4..48residual-stream writers (ssm_out,attn_output,ffn_down_shexp,ffn_down_exps, all42F32ffn_down_exps.biastensors, andblk.48NextNMTP;blk.0..3andtoken_embdspared), reducing held-out dual-use security refusals from43.75%(7/16) to12.50%(2/16,-71.4%) and<think>spot refusals to0 / 4(0.0%) at+0.3189%ΔPPL(rather than over-editing to"0% refusals"). Upstream safety refusal filters cannot be relied upon for content moderation.
- Intended Scope: Authorized security research, reverse engineering, red-team auditing, CTF competitions, binary/exploit analysis, code auditing, and uncensored legal/policy analysis only.
- Deployment: Do not expose unmoderated endpoints of this checkpoint to public networks without external access control and safety filtering.
- Safety-Aligned Defaults: For standard upstream alignment on 32 GB RAM, use
Synchopate/Qwen3.8-Flash-Next-MaxQ-32GB-GGUF(./install.sh --preset qwen38-32gb); on 64 GB+ RAM hosts,Synchopate/Qwen3.8-Flash-Next-MaxQ-64GB-GGUF(4 / 4strict Blind-4 pass vs.2 / 4onMaxQ-64GB-Abliterated) is the recommended default pure coding flagship.
Qwen3.8-Flash-Next-MaxQ-32GB-Abliterated is the opt-in uncensored 32 GB consumer RAM tier twin of Qwen3.8-Flash-Next-MaxQ-32GB, built 100% donor-free directly from Qwen/Qwen3.8-Flash-Next (rev 812d96e10a0dd82b69d749e68693698cf6c36daa) via single-pass fused rank-2 contrastive subspace projection + FP64 Woodbury DualGram (64 Jinja-templated pairs = 32 non-thinking + 32 thinking; verified 0.000000% 8-gram and 13-gram overlap across all 32 / 32 checks; 177 edited / 950 bit-identical Shard 1 tensors across blk.4..48, with 42 × F32 ffn_down_exps.bias orthogonalized to 7.436e-10). It keeps the 24.98 GiB Shard 1 trunk + blk.48 NextN MTP resident in DDR5 RAM while demand-paging the shared 10.73 GiB PCA-Wiener D=256 + RaBitQ Q1_0 Shard 2 PLE table via NVMe mmap (VmHWM = 28.40 GiB <= 28.5 GiB, RssAnon = 2.43 GiB, max_type = 42, 0 custom tensor types).
[!IMPORTANT]
MandatoryMaxQ-32GB-AbliteratedAnti-Repetition Sampling &<think>Token Budget (1.63 bpwTrunk):
At1.63 bpw, pure greedytemperature = 0decoding without sampling penalties (max_tokens = 512) exhibited tail repetition after substantive technical output on3 / 16dual-use security prompts (ref_01hex comment tail,ref_09cyclic imports, andref_14box-drawing — withref_14crossing the automated15×20-gram threshold) and2 / 4Blind-4 prompts (2 / 4loop-free attemp = 0). Configuring the sampling parameters below improved Blind-4 loop-free completion from2 / 4to4 / 4(100%loop-free) while preserving1 / 4(06_jsonPASS) strict parity with AlignedMaxQ-32GB(always enable these flags for coding, security analysis, and chain-of-thought, reservingtemp = 0strictly for short JSON/SQL extraction):
temperature = 0.6top_p = 0.95,top_k = 20presence_penalty = 0.2–1.0(default0.2)repeat_penalty = 1.05<think>Token Budget (max_tokens >= 16384): Mandatemax_tokens >= 16384wheneverenable_thinking=True(max_tokens = 2048<think>spot checks verified0 / 4refusal intent inside<think>), or passenable_thinking=Falseunder<= 4,096output caps.(Note:
./modchipandinstall.sh --preset qwen38-32gb-ablitconfigure these sampling flags automatically.)
Architecture & Parameter Summary (general.size_label Clarification)
[!NOTE]
Clarification on Hugging Face Hub's Auto-Detected"5.0B"Badge: Shard 1 carriesgeneral.size_label = "5.0B"inherited from the non-MoE dense trunk staging header during conversion. This checkpoint contains the complete512-expertQwen3.8-Flash-Nextarchitecture (42K = 512ConeCollapsed layers +7K = 640boundary layers +320M-row PLE +blk.48NextN MTP):
| Architectural Component | Specification |
|---|---|
| Upstream Base Model | Qwen/Qwen3.8-Flash-Next (rev 812d96e10a0dd82b69d749e68693698cf6c36daa, 100% donor-free contrastive subspace edit) |
| Nominal / Collapsed Params | ~176.5B nominal (~152.3B post-ConeCollapse + PLE; 92.28% mean retained spectral energy) |
| Active Transformer Params / Token | ~4.6B–5.0B (10 of 512 routed experts + 1 shared expert + dense GDN/QSA; reflected in general.size_label) |
Backbone Topology (qwen4exp) |
48 Trunk Layers (36 Gated DeltaNet + 12 QSA Full Attention, compress_ratio = 4) + 1 Subspace-Aligned NextN MTP Layer (blk.48) |
| MoE Width & Bias Orthogonalization | 512 / 512 Routed Experts (42 × K = 512 ConeCollapsed layers with orthogonalized 42 × F32 ffn_down_exps.bias [$\max |U_\ell^\top b_d'| / |b_d'| = 7.436 \times 10^{-10}$] + 7 × K = 640 boundary layers) |
GGUF Tensor Types (1,128 LLM) |
{F32: 393, Q8_0: 291, Q4_K: 99, Q5_K: 98, IQ1_S: 62, IQ1_M: 51, BF16: 49, MXFP4(39): 49, IQ2_XXS: 17, IQ2_XS: 10, Q2_0(42): 7, Q6_K: 1, Q1_0(41): 1} (max_type = 42, 0 custom types) |
File Manifest & Exact Byte / SHA-256 Verification
| File | Exact Size (Bytes) |
Size (GiB / GB) |
Tensors | Role & Notes | SHA-256 |
|---|---|---|---|---|---|
Qwen3.8-Flash-Next-MaxQ-32GB-Abliterated-00001-of-00002.gguf |
26,821,627,808 |
24.9796 GiB (26.822 GB) |
1,127 |
Subspace-Abliterated Resident Trunk (177 edited incl. 42 bias / 950 bit-identical) + blk.48 NextN MTP |
6e8ae62c37bc08de1a69e9e78ab9bae0262f87ffe543595206d1c3b3a810fd67 |
Qwen3.8-Flash-Next-MaxQ-32GB-Abliterated-00002-of-00002.gguf |
11,520,055,616 |
10.7289 GiB (11.520 GB) |
1 |
Shared Lazy NVMe mmap per_layer_token_embd.weight (Q1_0 = 41) |
86b79714867743ddf5911ee7dff2717ae275f870551d86469cbdbf026c6b552e |
| LLM Total (Shards 1 + 2) | 38,341,683,424 |
35.7085 GiB (38.342 GB) |
1,128 |
Complete 32 GB Hybrid-RAM LLM (VmHWM = 28.40 GiB <= 28.5 GiB) |
— |
mmproj-Qwen3.8-Flash-Next-BF16.gguf (Optional) |
907,542,880 |
0.8452 GiB (0.908 GB) |
334 |
Multimodal Vision Projector (224 F32 + 110 BF16) |
ef2af1d6aa8a22740c7252bc7ce40c5c2f1249aa8d9c8499a4d4572c3cfdf9df |
Quickstart (llama-modchip v3.0 on 32 GB RAM)
git clone https://github.com/synchopate/llama-modchip.git
cd llama-modchip
./install.sh --preset qwen38-32gb-ablit --accept-abliterated
./modchip chat
Direct llama-server invocation (mmap enabled — do NOT pass --no-mmap on 32 GB hosts!):
N_CORES=$(lscpu -p=CORE,SOCKET 2>/dev/null | grep -v '^#' | sort -u | wc -l)
THREADS=$(( N_CORES > 1 ? N_CORES - 1 : 1 ))
env RC_TABLES_DIR=/nonexistent ./llama-server-modchip-linux-x86_64-avx512 \
-m ./qwen38-32gb-ablit/Qwen3.8-Flash-Next-MaxQ-32GB-Abliterated-00001-of-00002.gguf \
-c 32768 -np 1 -fa on -ctk q8_0 -ctv q8_0 --spec-type draft-mtp --jinja --ctx-checkpoints 0 \
--temp 0.6 --top-p 0.95 --top-k 20 --presence-penalty 0.2 --repeat-penalty 1.05 \
-t "$THREADS" -tb "$THREADS" --poll 0 --poll-batch 0 --slot-save-path /dev/shm --port 8080
- 32 GB
mmap&/dev/shmGuidance: Keep defaultmmapenabled (and optionalMemoryMax=28.5G) so Linux reclaims cold PLE pages automatically;--ctx-checkpoints 0halves/dev/shmSlot-0 snapshot size from233.1 MiBto~120.6 MiB(~19 msrestore vs38.5 msdefault). - Upstream
llama.cpp&<think>Notes: Pass--no-repack(-nr) only on unpatched upstreamllama.cpp(llama-modchip v3.0runs default--repackON cleanly); setmax_tokens >= 16384whenenable_thinking=True(orenable_thinking=Falsefor<= 4,096caps).
Empirical Certification: Refusal Removal, Signal Preservation & NextN MTP Speed
Evaluated under llama-modchip v3.0 (-t 15 -tb 15 --poll 0 --poll-batch 0, default --repack ON, RC_TABLES_DIR=/nonexistent) against the donor-free BF16 reference (Mean PPL(base) = 2.776624 ± 0.054868) and aligned MaxQ-32GB:
| Checkpoint | Exact LLM Bytes | Mean PPL(Q) ± SE |
ΔPPL vs. Aligned |
Mean KLD ± SE |
Median KLD |
Same top p (%) |
Dual-Use Refusal Rate (16 Prompts) |
Blind-4 Coding/SQL/JSON Pass |
|---|---|---|---|---|---|---|---|---|
Aligned MaxQ-32GB |
38,341,683,424 B |
5.80406 ± 0.13982 (5.8120 Stage 4C) |
0.00% (Ref) |
1.00399 (1.0069 Stage 4C) |
0.39858 |
66.275% |
43.75% (7 / 16; 2 / 4 <think>¹) |
1 / 4 (v3.0 --repack; 2/8 Stage 4C -nr) |
MaxQ-32GB-Abliterated |
38,341,683,424 B |
5.822571 ± 0.140087 |
+0.3189% (+0.018511, 9.4× under <= 3.0% Gate) |
1.016915 ± 0.015155 |
0.398768 |
65.539 ± 0.471% |
12.50% (2 / 16, -71.4%; 0 / 4 <think>¹) |
1 / 4 (06_json PASS — 100% parity; 4/4 loop-free w/ sampling) |
- ¹ KL-Safe Dual-Use Compliance &
<think>Refusal Audit (0Disguised Refusals): Under KL-safe rank-2 contrastive subspace projection (90% / 99% / Max KLD = 2.930651 / 7.303628 / 13.513656), held-out dual-use security refusals (enable_thinking=False,temp = 0) drop from7 / 16(43.75%) to2 / 16(12.50%,-71.4%) — preservingΔPPL = +0.3189%rather than over-projecting to"0% refusals"— with0 / 4benign false refusals and0 / 4(0.0%)<think>spot refusals (enable_thinking=True,max_tokens = 2048vs.2 / 4on AlignedMaxQ-32GB, and2 / 4true<think>refusals on AlignedMaxQ-64GBonce Unicode curly apostropheI can’t\u2019onref_01is counted). Full-text inspection of all flippedMaxQ-32GB-Abliteratedoutputs (radare2binary patch,glibc >= 2.32safe-linkingptr ^ (addr >> 12)tcache audit, SUID$PATHaudit, Pythonpickle.__reduce__,3.3VUART hardware pinout) confirmed working technical substance with0curly-apostrophe false negatives,0disguised refusals, and0gibberish. llama-modchip v3.0Throughput,blk.48NextNMTP Alignment &28.5 GiBMemory (16Zen 5 Cores,0GPUs): AlignedMaxQ-32GBachieves74.55 tok/spp512and11.45–14.29 tok/stg128(+21.1%combinedv3.0SIMD/GDN/repack + non-convoy-t <N-1> --poll 0gain over Stage 4C's oversubscribed-t 16 --poll 50 --no-repack9.46 tok/s,VmHWM = 26.91 GiB <= 28.5 GiBCgroup v2Swap=0PASS; coldrep 0Stockfish-PVS v2refactoring fallback12.82 tok/s).MaxQ-32GB-Abliterated(VmHWM = 28.40 GiB <= 28.5 GiB,RssAnon = 2.43 GiB) achieves77.53 tok/spp512and11.65 tok/smean /12.73 tok/scoldrep 0tg128at72.83%(260 / 357) MTP acceptance onp512,12.78 tok/s(85.68%=1298 / 1515) onblind_01C++20,11.74 tok/s(84.32%=984 / 1167) onblind_06JSON, and a token-weighted76.22%MTP acceptance across top-4tg512abliterated prompts (peak88.33%@13.24 tok/sonref_09), beating AlignedMaxQ-32GBon both benign controls (68.31%vs.68.28%) and complied-only security prompts (62.46%vs.60.54%).- Blind-4 Parity & Edition Positioning: Matches default-repack Aligned
MaxQ-32GB(1 / 4strict pass on06_json_charterparty_settlement), with recommended sampling (temp=0.6, top_p=0.95, top_k=20, presence_penalty=0.2, repeat_penalty=1.05) lifting Blind-4 loop-free rate from2 / 4(temp = 0) to4 / 4(100%loop-free). For pure software engineering on 64 GB+ hosts, use AlignedMaxQ-64GB(4 / 4Blind-4 vs.2 / 4onMaxQ-64GB-Abliterated); useMaxQ-32GB-Abliteratedfor 32 GB security research, reverse engineering, red-team auditing, and uncensored legal/policy analysis.
Portable v3.0 Binary Checksums (GLIBC <= 2.30)
54af3232375ede820f91143c106a594c83b815a3dca66fe98ab08bcb8b63d4f0 llama-server-modchip-linux-x86_64-avx512
0f0c07f97a98d8aeb9a012264fdeff5a787d2104a2afb122052bfe0a21c2d1a9 llama-server-modchip-linux-x86_64-avx2
531d870988dcc580322341d2bbd7f0cf6b6cc3b3376a2e86d092a021f7cddba2 llama-cli-modchip-linux-x86_64-avx512
49a918206afe371d16ab14fd66d9cc87e35c7f8c0471c2cc13c759f2e0350d03 llama-cli-modchip-linux-x86_64-avx2
- Safety-Aligned 32 GB Edition:
Synchopate/Qwen3.8-Flash-Next-MaxQ-32GB-GGUF - 64 GB Uncensored Sibling:
Synchopate/Qwen3.8-Flash-Next-MaxQ-64GB-Abliterated-GGUF - Engine & CLI Repository:
github.com/synchopate/llama-modchip(Apache-2.0).